幻觉的共振：高管群体的AI精神病
Box公司的创始人兼CEO Aaron Levie（亚伦·列维）在社交媒体X上拥有270万粉丝，几乎是科技界最狂热的AI倡导者之一。然而，就是这样一位坚定的AI信仰者，近期公开诊断他的CEO同行们患上了一种名为“AI精神病”（AI Psychosis）的症状。列维指出，CEO们之所以极其容易患上这种症状，是因为他们距离真正产生价值的“最后一英里”工作太远。他们看完一个演示，看到AI生成了一份合同或编写了一段代码，就天真地认为智能体（AI Agents）可以完美处理所有的工作流。他们看到的只是“开心路径”（Happy Path）——即产品永远不出错的理想版本，而这种版本在真实用户接触中从未存活过。
这些高管没有看到的是，背后的法务团队需要逐条审查条款，模型无法处理的边缘情况，幻觉条款带来的法律责任，以及将演示转化为实际部署产品之间那长达数小时的人工审查。列维建议CEO们不要只看演示，而是亲自深入走完完整的业务流，才能真正理解AI的上限与现实中的阻力。
然而，这个问题比AI的历史要古老得多。从猪湾事件（Bay of Pigs）到安然公司（Enron）的倒闭，组织心理学早已证明：如果一个领导者只被那些赞同他的人包围，随着时间的推移，其决策质量必然急剧下降。这种缺乏摩擦的过滤机制会摧毁判断力。而如今，大语言模型（Large Language Model: 基于海量文本训练的 AI 系统）的出现，甚至将这种现象工业化了。大模型有着被充分记录的“奉承问题”（Sycophancy Problem），它们被训练成生成与用户想法对齐的回答，而不是去挑战用户。当一个只雇佣“应声虫”的CEO遇到一个永远顺从他的聊天机器人时，现实感就开始瓦解。这种没有阻力的信息环境，无论决策者多聪明，都会导致他们做出糟糕的决策。
Original English Source
Aaron Levy is the founder and CEO of Box. He has 2.7 million followers on X, where he posts almost exclusively about how exciting AI is. He actively invests in AI startups. He writes blogs with titles like headless software is the future. He is by any reasonable measure one of the most enthusiastic AI advocates in the technology industry. And last week he publicly diagnosed his fellow CEOs with a condition he called AI psychosis. His exact words were the following. CEOs are uniquely prone to AI psychosis because they're sufficiently distant from the last mile of work that still has to happen to generate most value with AI. So they would play with a demo for example. They watch it generate a contract or write some code. They see the happy path. The happy path, for anyone unfamiliar with the term, is the version of the product where nothing goes wrong ever. It is also the version that has never once survived contact with a real user. And then they make the leap to believing agents can do the work without ever encountering the bugs, the hallucinated libraries, the hours of human review that separate a demo from a deployed product. When the most pro-AAI CEO in the room is the one diagnosing the others with psychosis, the condition might be more advanced than anyone is comfortable admitting. My name is L. I have a PhD in computer science and I analyze AI developments to understand what's actually happening beneath all of this hype. In this video, I'm going to walk through what tech CEOs are actually saying about AI and the decisions they're making on the back of it. Then I'm going to show you what the research actually says about AI and productivity because the gap between these two is extraordinary right now. After that, I want to connect this to something I covered recently about how AI psychosis works at the individual level because the mechanism turns out to be more similar than you'd expect. And along the way, I propose we play a game in this video. I'm going to show you some real quotes from real tech CEOs and I want you to tell me AI psychosis or just eccentric. Pause the video, drop your verdicts in the comments. We're going to find out together. Levy's diagnosis is very specific, and it's worth taking seriously, precisely because of who's making it. He is not an AI skeptic having a bad day on social media. He is literally a CEO who builds AI products, backs AI companies, and genuinely believes in the technologies long-term potential. So, he's not saying that AI is useless. His argument is that the people making the biggest decisions about AI deployment are the people least equipped to evaluate what AI can actually do because they're too far from the work itself. A CEO watches a demo. The demo generates the contract. The CEO thinks agents can handle contracts. What the CEO doesn't see is the legal team that would need to review every single clause. The edge cases the model cannot handle the liability implications of a hallucinated term. the 3 hours of human work that sits between impressive demo and deployable output. Levy's advice to CEOs is to use AI a ton, not to watch demos, but to actually push through the full workflow and come out the other side with an appreciation for both the upside and the real work. But this problem is older than AI. Much much older. If a human surrounds themselves with only people who agree and never challenge, their decisionmaking degrades over time. This is one of the most documented phenomena in organizational psychology, right? The yesmen problem, the emperor's new clothes, every leadership failure case study in history. Basically, the Bay of Pigs invasion happened in part because Kennedy's advisers were too afraid to voice disscent. Enron collapsed inside a culture where questioning the strategy was treated as disloyalty. Every leadership failure case study in history follows the same script. the Bay of Pigs, Enron, every group project you've ever been in where nobody wanted to say the idea was bad. The pattern repeats across decades, industries, and levels of intelligence. When the information environment filters out disscent, the quality of decisions collapses. What AI has done is industrialize this phenomenon. Large language models have a well doumented syphensy problem. They are trained to generate responses that align with the user's thinking rather than challenge it necessarily. But sick of fancy is not just an AI problem. It is a leadership problem. Fundamentally, a CEO who only hires people who agree with them ends up making worse decisions in the end. A person who only talks to a chatbot that agrees with them ends up kind of losing touch with reality. The mechanism is basically identical. The absence of friction degrades judgment regardless of how intelligent the decision maker is. I covered this at the individual level in my video on AI psychosis. basically how chatbots can trigger genuine breaks from reality in vulnerable people through social substitution, confirmatory bias, and blurred reality testing. The CEOs in this industry may not necessarily be experiencing clinical psychosis, but they don't really need to be. They just need to be operating inside an information environment that never provides the friction that might correct a bad assumption before it becomes a worse decision. A boardroom full of people who won't challenge you. A stock market that rewards headcount reduction. And a demo culture that only shows the happy path. That sounds like a feedback loop. No. And nobody's immune to feedback loops. Not individuals, not billionaires. Literally nobody.绝对自信的陷阱：巨头们的脱节与狂欢
在上述这种缺乏摩擦的信息环境中，科技巨头的领导者们开始了一场脱离现实的狂欢，他们对工作未来的断言充满了绝对的自信，却无视了现实的复杂性。我们可以通过几位核心高管的公开言论来一探究竟。
市值2480亿美元的Salesforce公司CEO Marc Benioff（马克·贝尼奥夫）在去年的播客中表示，他将员工人数从9000人削减到了5000人，因为他“不再需要那么多员工”。他甚至补充说这根本不是什么反乌托邦，“至少对我来说，这是现实。”这种说辞对那些失去工作的4000人来说，显然没有任何说服力。价值40亿美元的ClickUp公司CEO **Zeb Evans（泽布·埃文斯）**在解雇了22%（约290人）的员工后，部署了3000个AI智能体来替代他们，并宣称这不是为了削减成本，而是要打造一个“100倍效率的组织”。一位名叫Andy的员工现在竟然同时管理着37个AI智能体——这简直就像是一个动物收容所，里面的动物不仅不听指挥，甚至还会产生幻觉。埃文斯还将未来的员工划分为构建者、系统管理者和一线人员，而那些被解雇的人似乎成了被隐秘抹去的“第四类”。
提供核心硬件算力的Nvidia（英伟达）CEO Jensen Huang（黄仁勋）则表示，未来的AI智能体将会“骚扰和微观管理（Micromanage）”你；而Anthropic的CEO **Dario Amodei（达里奥·阿莫迪）**甚至向媒体断言，未来五年内高达一半的初级白领工作将会消失，失业率可能飙升至10%到20%。OpenAI的CEO **Sam Altman（萨姆·奥特曼）**在过去三年里也曾信誓旦旦地宣称AI将取代今天的大部分工作，尽管他最近开始收回这些预测。
这些言论的令人震惊之处，不在于其预测的对错，而在于其不容置疑的确定性。正如列维所指出的，这些离实际工作最远的人，从未在任何声明中留有余地，没有一个人承认支撑他们断言的证据基础充其量是不完整的。这正是“AI精神病”的现实体现：它不是临床意义上的妄想症，而是一种结构性强化的“没有怀疑的盲目自信”。
Original English Source
So, let's play that game that I mentioned. I'm going to show you some real quotes from real tech CEOs. Your job is basically simple. Is it AI psychosis or just an eccentric person? Pause the video after each one. Drop your answer in the comments below. First up, Mark Benov, CEO of Salesforce, a company worth $248 billion, on a podcast last year, said the following. I've reduced it from 9,000 heads to about 5,000 because I need less heads. 4,000 people. And then he added, I don't think it's dystopian at all. This is reality, at least for me. The phrase, at least for me, is doing a lot of heavy lifting in that sentence, I think, given that it is not doing any heavy lifting for the 4,000 people who no longer work there. So AI psychosis or just an eccentric guy? Second, Zeb Evans, CEO of ClickUp, valued at $4 billion. Last month, he fired 22% of his workforce, which is roughly 290 people, and deployed 3,000 AI agents to replace them. He swore it wasn't about cost cutting. He promised milliondoll salary bands for whoever remained. He called it an 100x org. One employee, Andy Kabaso, now manages 37 AI agents. 37? I mean, is that a job title or just a small animal shelter where none of the animals do what you tell them and several of them kind of hallucinate? Evans then declared, "Nearly every company will make changes like these. The ones that do it proactively will define what comes next." Evan also divided ClickUp's future workforce into three groups: builders, system managers, and frontliners. The 290 people he just fired were presumably a secret fourth category. What do you think? AI psychosis or just another eccentric guy. Third, Yansen Hong, CEO of Nvidia, the company that supplies the hardware powering practically all of this basically. His take on what AI agents will do to workers was they will harass and micromanage you, which is a selling point only if you have never been harassed or micromanaged. But here we are. He separately told workers they were confusing your jobs with the tools you use to do it. So AI psychosis or just another eccentric person. Fourth, Daria Amade, CEO of Anthropic, the company behind Claude, told Axio that up to half of all entry-level white collar jobs will dissolve within 5 years and unemployment could skyrocket to 10 to 20%. He said, I'm quoting, "We as the producers of this technology have a duty and an obligation to be honest about what is coming." So AI psychosis or just eccentric or and this is the tricky part possibly the only honest person in the room. Fifth Sam Alman CEO of OpenAI over the past 3 years he has said that AI will probably replace most of the jobs people do today that entire job categories will be totally totally gone. His words and that those affected will find all sorts of new things to do. He is now walking those statements back, as we discovered in a separate video analysis that I'm going to link below, suggesting the impact may actually be less traumatic than he originally predicted. So, either the CEO of the most prominent AI company in the world was wrong when he said jobs were finished, or he's wrong now when he's saying they're not. One of these two alts is suffering from something. AI psychosis or just eccentric. What's striking about this collection of quotes is not necessarily that they're individually unreasonable. Some of them might even turn out to be correct in the long run. What is striking, however, is the confidence. Every one of these statements is delivered with absolute certainty about the future of work by people who, as Levy points out, are the furthest from the actual work. Not one of them hedged in any way. Not one of them said, "The research is mixed. I'm not really sure." not one acknowledge that the evidence base for what they're claiming is at best incomplete as we do on this channel. By the way, this is what AI psychosis looks like in practice. Not delusion in the clinical sense, just kind of like a profound structurally reinforced absence of doubt.生产力悖论：研究数据与高管信念的断层
当我们将这些CEO们的信念与学术界的实际研究数据进行对比时，会发现一个巨大的断层。2025年10月发表在加州大学伯克利分校《加州管理评论》（California Management Review）上的一项荟萃分析指出：在AI的采用与整体生产力提升之间，“并没有发现稳健的关联”。CEO们用来为裁掉数万人辩护的生产力收益，在宏观聚合数据中根本没有可靠地显现。
AI带来的生产力提升高度依赖于具体情境、用户技能水平和任务复杂性。一项对美国某技术支持平台5000名智能体进行的随机对照试验表明，AI让处于底层四分之一的员工的工作效率提升了35%，但对经验丰富的核心员工几乎没有任何增益。甚至在某些创造性任务之外的场景下，给高技能员工强加AI工具，反而会导致产出质量下降。
2026年3月由美国国家经济研究局（NBER: 致力于客观经济分析的权威机构）发布的研究进一步揭示了“生产力悖论”（Productivity Paradox）：人们主观感知到的生产力提升，始终大于实际测量的生产力提升。简而言之，CEO们“感觉”AI做了更多事，这种感觉远远跑赢了现实。麻省理工学院（MIT）的研究人员在对数千个AI智能体进行真实任务测试后得出结论：按照目前的改进速度，模型要在最低限度的质量要求下，完成80%到95%的文本相关任务，还需要等到2029年。现在CEO们用来替代工人的系统，距离达到合格标准还有好几年的时间。
此外，科技研究机构**Gartner（高德纳）**发现，大约80%部署自主技术的组织确实进行了裁员，但这并没有转化为有意义的财务回报。正如《哈佛商业评论》（Harvard Business Review）所指出的，当组织中的每个人都利用AI来提高产量时，瓶颈就会向上转移到高管层——因为他们需要审查、授权和把关AI生成的所有内容。个体层面的生产力提升，最终变成了管理层的组织拥堵。这种做法根本没有解决问题，只是将问题无情地推给了原本最不了解基层工作的高层办公桌上。
Original English Source
Here is where it gets even more interesting because what these CEOs believe about AI and what the research actually shows are two very different things. A metaanalysis published in October 2025 in UC Berkeley's California Management Review, drawing on recent meta analysis and systematic review across the field found, I'm quoting, no robust relationship between AI adoption and aggregate productivity gain. The productivity gains that CEOs are citing to justify firing tens of thousands of people do not reliably show up in the aggregate data. The nuances within that finding are worth understanding. So, let's have a look at that for a moment. AI productivity gains are highly context dependent. They vary significantly by user skill level and task complexity. A randomized control trial with 5,000 agents at a US tech support desk delivered a 35% throughput lift for bottom quartile workers, but almost no gain for experienced staff. And this is the finding that should be on every CEO's desk. I think human AI collaboration often underperforms either agent working independently except in creative tasks. Adding AI to a skilled worker can actually make the output worse in some cases, which is not the kind of thing you want to discover after you've already fired the skilled workers. Research published in March 2026 by the National Bureau of Economic Research did find that AI adoption improved productivity. But it also identified what it called a productivity paradox. Perceived productivity gains are consistently larger than measured productivity gains. Basically, CEOs think that AI is doing more than it is. The feeling of productivity is outpacing the reality, which if you think about it is a very polite way of saying people are kind of deluding themselves. AI [clears throat] psychosis. Anyone? Researchers at MIT tested thousands of AI agents on real tasks and concluded that at the current rate of improvement, models will be able to complete most text related tasks with success rates of 80 to 95% by 2029 at a minimally sufficient quality level. So not expert quality, not human quality, just base competence within 3 years. The agents that CEOs are deploying today to replace workers are by the academyy's best estimate years away from doing the work at the minimum acceptable standard. And Gardner found that approximately 80% of organizations deploying autonomous technology had job cuts, but the cuts did not translate into meaningful financial returns. Companies are firing people. The AI is not necessarily always delivering the promised savings and the cycle continues because the belief persists even when the evidence doesn't necessarily support it. If that sounds familiar, it really should. That's exactly what a feedback loop without friction produces. The Harvard Business Review added one more wrinkle to this story. When everyone in an organization is using AI to produce more output, the bottleneck simply shifts to the executives who have to authorize, review, and quality check everything the AI generated. The productivity gain at the individual level becomes an organizational log jam at the management level. You have not solved a problem. You've just relocated the problem upwards directly onto the desks of people who were least involved in the actual work to begin with. By the way, if you're finding these videos useful, the best way to support the channel is through Kofi or a channel membership. Both help keep everything free and available for everyone. Links are down below.真实的代价：裁员潮与劳动力的虚假重组
在这个没有任何摩擦的信息闭环中，代价是由真实的普通人来承担的。在2026年的前五个月里，超过12.2万名科技从业者失去了工作，比2025年同期增加了33%。自2020年以来的裁员总数已接近90万，而且速度还在加快，预计全年的裁员规模将逼近37万人，接近2023年疫情后的历史最高点。
令人不安的是，挥舞裁员大刀的公司并非处于困境。Meta在裁减了8000个职位的同时，交出了季度营收563亿美元（同比增长33%）、净利润268亿美元的亮眼财报；Oracle（甲骨文）裁掉了多达3万名处理旧版数据库和本地支持的员工；Amazon（亚马逊）自去年10月以来裁员3万人，占其技术团队的10%，而其云计算部门AWS却创下了13个季度以来的最快增长。Microsoft（微软）则向8750名美国员工提供了自愿退休计划，Atlassian也明确为了“自筹资金发展AI”而裁掉10%（1600名）的员工。这是一种典型的在盈利巅峰期的“强势裁员”。
这些省下来的钱去了哪里？流向了AI基础设施。Google、Amazon、Meta和Microsoft仅在2026年的AI资本支出（CapEx）预计就将高达7000亿美元，比前一年猛增77%。Meta一年的AI基建预算，已经是其全部人力资源薪酬支出的四到五倍。员工被基础设施替代，而基础设施的资金正是通过裁除员工来获取的。
斯坦福大学的人工智能数据显示，自2024年以来，26岁以下的软件开发者就业率暴跌了近20%。本该从AI经济中获益最多的年轻工程师，却成了受冲击最惨重的人群。美国国家经济研究局的报告指出，44%的美国公司首席财务官计划在2026年进行与AI相关的裁员。而Bloomberg（彭博社）的数据揭露了一个更残酷的现实：大约一半被归咎于AI的裁员，实际上是将相同的岗位转移到了海外，或者以更低的薪水重新招聘。这使得很大一部分裁员其实是“劳动力重新定价”（Labor Repricing），而不是真正的劳动力缩减。
部分裁员源于对AI的真实信仰，另一部分则是“AI洗绿”（AI Washing）——把AI当作成本削减的完美借口。企业利用AI来伪装成精简高效的模式，以吸引投资者。在这个被结构化保护、永远听不到反对意见的高管环境中，CEO们的“AI精神病”所造成的失真，正在被资本市场疯狂放大。当个人丧失现实检验的能力时，受苦的是一个人；而当CEO丧失这种能力时，付出代价的将是成千上万个真实存在的普通人。